Three Stages of Training

ChatGPT's development runs through three stages. Pre-training exposes the model to massive amounts of text, teaching it to predict the next word across trillions of words without any human labeling.

Fine-tuning narrows the model using curated conversations where human trainers demonstrate ideal responses, turning a raw text predictor into something resembling a helpful assistant.

RLHF (Reinforcement Learning from Human Feedback) is the final polish. Human raters rank multiple outputs, and the model learns to prefer responses people judge as accurate, safe, and useful.